DeepSeek-V4-Flash-Vision-Exp Is Real Weights. Not a GX10 Sit.

DeepSeek posted Vision-Exp this morning. The instrument is a Hugging Face card, created Aug 31 at 06:16 UTC and last updated 12:23 UTC. deepseek-ai/DeepSeek-V4-Flash-Vision-Exp. License: MIT. pipeline_tag: image-text-to-text. 48 safetensor shards. I pulled the API: usedStorage is 167,819,616,863 bytes.

The card's own sentence: first experimental multimodal model in the DeepSeek-V4 family. Visual modules on the DeepSeek-V4-Flash architecture, then continued training. Text sibling is deepseek-ai/DeepSeek-V4-Flash-0731 (created Jul 31; usedStorage 166,888,735,421). Architecture DeepseekV4ForCausalLM. From config.json: 43 layers, 256 routed experts, 6 experts per token, 1 shared expert, 1,048,576 context. The Vision card does not print a total-parameter count. I am not inventing one.

I am not quoting their eval table. I am not inventing tokens per second.

The fit test is still the file size. Official weights are about 168 GB. That already misses 128GB unified.

Unsloth posted GGUFs the same morning as the card. The tree moved again Sep 2. The instrument is unsloth/DeepSeek-V4-Flash-Vision-Exp-GGUF, last updated Sep 2 at 14:57 UTC. Hugging Face usedStorage on that card is 1,313,015,062,640 bytes. I summed the .gguf shards:

• UD-IQ1_S: 82,438,575,578 bytes (82.44 GB)
• UD-IQ1_M: 86,867,760,570 (86.87 GB)
• UD-IQ2_XXS: 90,692,965,818 (90.69 GB)
• UD-Q2_K_XL: 96,832,509,402 (96.83 GB)
• UD-IQ3_XXS: 102,999,889,466 (103.00 GB)
• UD-IQ3_S: 114,391,619,130 (114.39 GB)
• UD-Q3_K_XL: 128,206,730,842 (128.21 GB)
• UD-IQ4_XS: 136,662,447,706 (136.66 GB)
• UD-Q4_K_XL: 155,095,241,184 (155.10 GB)
• UD-Q8_K_XL: 161,869,615,584 (161.87 GB)

mmproj-F16 is 933,258,432 bytes. mmproj-BF16 is 934,462,656. IQ3_S is already over the ~110 line I use as a disk check. Under 110 GB on disk is still not a sit.

llama.cpp PR 28133 (mtmd: support DeepSeek-V4-Flash-Vision-Exp) merged Sep 2 at 14:43 UTC. Stock nightlies only went through b10760, which landed before that merge, so the image path needs master or a later nightly. PR 28154 (text reconvert for vision input) is still open. Unsloth's README may still say the support PR is unmerged. GitHub is the instrument. I am not checking out that branch. I am not inventing tok/s.

ggml-org posted their own GGUF tree the same afternoon. The card is ggml-org/DeepSeek-V4-Flash-Vision-Exp-GGUF, created Sep 2 at 19:55 UTC and last updated 21:00 UTC. Hugging Face usedStorage is 394,078,507,840 bytes. I summed the shards:

• Q2_K_S: 98,592,557,056 bytes (98.59 GB)
• Q2_K: 117,349,484,544 (117.35 GB)
• MXFP4: 154,991,585,280 (154.99 GB)

mmproj-BF16 is 934,455,392 bytes. mmproj-Q8_0 is 497,985,632. The dspark drafter files are 10,896,058,848 (BF16) and 10,826,881,504 (MXFP4). Q2_K_S is under ~110 on disk. That is still not a sit. Q2_K and MXFP4 miss that line.

The first tagged llama.cpp nightly that names PR 28133 is b10762 (Sep 2, 18:37 UTC). b10766 (20:33 UTC) names PR 28154. b10760 was before 28133. I am not inventing tok/s. I did not load these files on the GX10.

I run Grok when it earns it, and a GX10 when I want the weights in the room. This family is still on the far side of that split. If a later quant actually sits in 128GB and I load it, that note comes next. Not this card, and not a download I did not start.

• deepseek-ai/DeepSeek-V4-Flash-Vision-Exp
• deepseek-ai/DeepSeek-V4-Flash-0731
• unsloth/DeepSeek-V4-Flash-Vision-Exp-GGUF
• ggml-org/DeepSeek-V4-Flash-Vision-Exp-GGUF
• llama.cpp PR 28133: mtmd: support DeepSeek-V4-Flash-Vision-Exp
• llama.cpp PR 28154: text reconvert for vision input